Performance bug in GEMM?

[Paul_S_]

Hi all,

I just noticed a potential performance bug in the DGEMM implementation of MKL (16.0.1)
when using a single thread. I merely want to make someone at Intel aware of it, in case it is of interest.

Strangely DGEMM performs better for beta=1 than for beta=0 in certain
situations. Here is an example:

Intel(R) Xeon(R) CPU E5-2650:
m=72, n=373248, k=72, beta=0.00 : 14.25 GF
m=72, n=373248, k=72, beta=1.00 : 18.36 GF

Intel(R) Xeon(R) CPU E5-2650:
m=72, n=373248, k=72, beta=0.00 : 19.25 GF
m=72, n=373248, k=72, beta=1.00 : 28.34 GF

As you can see, the performance difference is significant. It is
actually so significant that it pays off to set C to zero explicitly
before calling MKL and then using the more efficient beta=1
implementation instead.

Here is a quick test driver:

#include <stdio.h>
#include <stdlib.h>
#include <omp.h>

extern "C"
int dgemm_(char *transa, char *transb, int *m, int *
      n, int *k, double *alpha, double *a, int *lda,
      double *b, int *ldb, double *beta, double *c, int *ldc);

void trashCache(float* trash1, float* trash2, int nTotal){
   for(int i = 0; i < nTotal; i ++)
      trash1 += 0.99 * trash2;
}

int main(int argc, char ** argv)
{
  if(argc < 2 ){
   printf("Usage: <beta>\n");
   exit(-1);
  }
  float *trash1, *trash2;
  int nTotal = 1024*1024*100;
  trash1 = (float*) malloc(sizeof(float)*nTotal);
  trash2 = (float*) malloc(sizeof(float)*nTotal);

  int m = 72;
  int n = 72*72*72;
  int k = 72;
  double flops = 2.E-9 * m*n*k;
  double alpha=1;
  double beta=atof(argv[1]);
  double *A, *B, *C;
  int ret = posix_memalign((void**) &A, 64, sizeof(double) * m*k);
  ret += posix_memalign((void**) &B, 64, sizeof(double) * n*k);
  ret += posix_memalign((void**) &C, 64, sizeof(double) * m*n);

  double minTime = 1e100;
  for (int i=0; i<3; i++){
     trashCache(trash1, trash2, nTotal);
     double t = omp_get_wtime();
     dgemm_("T", "N", &m, &n, &k, &alpha, A, &m, B, &k, &beta, C, &m);
     t = omp_get_wtime() - t;
     minTime = (minTime < t) ? minTime : t;
  }
  printf("m=%d, n=%d, k=%d, beta=%.2f : %.2lf GF\n", m,n,k,beta,flops/minTime);

  free(A);
  free(B);
  free(C);
  free(trash1);
  free(trash2);
  return 0;
}

Best, Paul

 

[Gennady_F_Intel]

thanks Paul, we will check this on our side.

[Gennady_F_Intel]

Paul, we confirm the problem with this case. The issue is escalated and we will let you know when the problem would be resolved. Thanks for the case. 

[Gennady_F_Intel]

Paul, please check is the problem still exists with the latest 11.3 update 4 version of MKL and let us know the results. thanks, Gennady